Derivation of multivariate syndromic outcome metrics for consistent testing across multiple models of cervical spinal cord injury in rats

Abstract

Spinal cord injury (SCI) and other neurological disorders involve complex biological and functional changes. Well-characterized preclinical models provide a powerful tool for understanding mechanisms of disease; however managing information produced by experimental models represents a significant challenge for translating findings across research projects and presents a substantial hurdle for translation of novel therapies to humans. In the present work we demonstrate a novel 'syndromic' information-processing approach for capitalizing on heterogeneous data from diverse preclinical models of SCI to discover translational outcomes for therapeutic testing. We first built a large, detailed repository of preclinical outcome data from 10 years of basic research on cervical SCI in rats, and then applied multivariate pattern detection techniques to extract features that are conserved across different injury models. We then applied this translational knowledge to derive a data-driven multivariate metric that provides a common 'ruler' for comparisons of outcomes across different types of injury (NYU/MASCIS weight drop injuries, Infinite Horizons (IH) injuries, and hemisection injuries). The findings revealed that each individual endpoint provides a different view of the SCI syndrome, and that considering any single outcome measure in isolation provides a misleading, incomplete view of the SCI syndrome. This limitation was overcome by taking a novel multivariate integrative approach for leveraging complex data from preclinical models of neurological disease to identify therapies that target multiple outcomes. We suggest that applying this syndromic approach provides a roadmap for translating therapies for SCI and other complex neurological diseases.

Figure 1. Standardized observation-based behavioral batteries for evaluating recovery after spinal cord injury in the rodents.

A, Grooming scale scoring system and recovery plots color coded by injury conditions. B, Paw placement task and recovery plots. C, Basso, Beattie, Bresnahan (BBB) open field hindlimb locomotor scale. D, Fine motor-focused, BBB subscore. E, Forelimb open field score. Three different standardized models of SCI were included in the dataset: hemisections, force-driven contusions (kdyns) and weight-drop contusions (mm) centered at cervical vertebra 5 (C5) and delivered to one side of the spinal cord. Data were collected over 10 years at two different SCI centers (The Ohio State University, and University of California, San Francisco) and represent over 159 subjects with complete outcome batteries. Error bars reflect SEM as used by general linear models (e.g., ANOVA). Note that all points have error bars although some are smaller than the points. (see manuscript for references).

Figure 2. Standardized digital locomotor analysis for evaluating recovery after spinal cord injury in rodents.

A, Digital footprint analysis allows objective quantification of many correlated outcomes including: B, Stride-length for each limb; C, Print area for each limb; D, Distribution of limb use reflected as the absolute deviation from the pre-injury baseline (i.e. deviation from ∼25% recruitment for each limb).

Figure 3. Histological outcomes.

A, Tissue sparing measures in SCI research are typically taken at the lesion center as determined by the largest extent of the lesion ellipsoid. Although specific methods for quantification may vary across studies, typical measures include lesion size, B, gray matter (GM) sparing, C, white matter (WM) sparing, D, total sparing (GM+WM), E, total tissue area (GM+WM+debris), F, motorneuron number. Scale bar, 100 µm. Since the compiled dataset was limited to unilateral injuries (hemisections or hemicontusions), all measures are represented as a percentage of the contralateral, spared hemicord. The quantified area is illustrated in red on a representative example. The representative example was taken from the subject closest to the group mean for lesion size across the study’s 159 subjects.

Figure 4. Multivariate analysis of the SCI syndrome using data from two research sites.

A, Heat map of the bivariate correlation matrix, indicating all cross-correlations between behavioral and histological outcomes sorted in a randomized fashion. Blue indicates negative relationships and red indicates positive relationships. Heat reflects magnitude of Pearson correlation (r). B, Zoomed view of a small portion of the correlation matrix showing the interrelationships between a subset of outcomes. C, Principal components analysis (PCA) by eigenvalue decomposition was used to reduce the correlation matrix to synthetic multivariate variables known as principal components (PCs). PCs reflect clustered variance shared by numerous outcome measures. PC identities are indicated by significant PC loadings (arrows, loadings |>.40|). Each loading is equivalent to a Pearson correlation between individual outcomes and the PC. Loading magnitude is indicated by arrow width and heat (blue reflects negative and red reflects positive relationships). Exact loading values are shown next to each arrow. See Fig. S1 for non-significant loadings. D, Plot of individual subjects (N = 159) in the 3D multivariate syndrome space described by PC1-3. E–G, 2D plots of PC1-3 on their own axes. Significant differences: E,*P<.05 from sham, ** P<.05 from 75 kdyn and sham, §P<.05 from all groups except 6.25 mm. F, *P<.05 from sham, **P<.05 from all groups but sham, ***P<.05 from sham, 75 kdyn, 100 kdyn and hemisection. §P<.05 from all other groups. G, *P<.05 from sham, ** P<.05 from 75 and 100 kdyn.

Figure 5. Consistency of PCA across subsets of variables.

A, Independent PCA extraction using only histological variables demonstrated significant replication of PC1 extracted using the full variable set. B, Injury condition affected PC1_HISTO in an equivalent manner to the full-variable extraction (compare to Figure 4E). C, PCA extraction using only behavioral variables significantly replicated the full variable PC for PC1 and PC2, however the sequence of extraction reversed, indicating a reversal in variance explained by PC1_BEHAVIOR and PC2_BEHAVIOR. D, Scores of individual subjects on PC1-3 extracted from just-behavioral variables. The pattern for PC2 _BEHAVIOR recapitulated PC1 from the full variable extraction (compare to Fig. 4E) and PC1_BEHAVIOR recapitulated PC2 from the full-variable extraction (compare to Fig. 4F). *P<.05 for replication statistics, s>0.63.

Figure 6. Consistency of multivariate syndromic patterns across two different biomechanically controlled cervical spinal contusion models.

A, SCI syndromic space extracted from an NYU/MASCIS injury device dataset (N = 52 rats; 24 outcome variables). B, SCI syndromic space extracted from an Infinite Horizons injury device dataset (N = 100 rats, 24 outcome variables). Note, normed PC score axes are scaled according to variance within each extraction, resulting in axes with units that are not directly comparable across extractions. However relative relationships among groups (sham vs. injuries) are conserved. C, Consensus PC loading patterns that are conserved across injury patterns. Loading weights (arrows) reflect average values across the two datasets. D, Statistical evaluation of PC cross-validation in the PC loading matrices from NYU/MASCIS and IH injury datasets. *p<.05 for n = 24 variables; s>0.63.